Generative biological foundation model placing DNA, RNA, and protein in one shared vocabulary, spanning genomic, proteomic, and cross-molecule tasks.
No providers recorded yet. Browse all providers
Biological sequence models are usually built one molecule at a time. Genomic language models learn nucleotides, protein language models learn amino acids, and connecting the two — asking which codons produce a protein, or whether an RNA binds a given polypeptide — means bolting separate encoders together and hoping their embedding spaces can be reconciled. The central dogma says these sequences are three views of the same information, but most architectures cannot express that.
CD-GPT (Central Dogma Generative Pretrained Transformer), from the Tencent AI for Life Sciences Lab, takes the opposite approach: a single decoder-only transformer with one billion parameters and one shared vocabulary covering nucleotide and amino-acid tokens alike. Because DNA, RNA, and protein occupy the same representational space, the model can be prompted with any of them and asked to predict or generate any other. It handles mono-molecular tasks such as promoter detection and protein solubility prediction, and genuinely cross-molecular ones such as RNA-protein interaction identification and reverse translation, where a protein sequence is decoded back into the codons that could encode it.
The model was released in June 2024 with a revised preprint in February 2025, and sits alongside single-substrate foundation models such as ESM-2 as an argument that multi-omic analysis benefits from a unified sequence space rather than a pipeline of specialists.
CD-GPT is a one-billion-parameter generative transformer pretrained on large-scale nucleotide and amino-acid corpora. The first stage trains on individual molecule types in isolation; the second trains on paired sequences drawn from the central dogma, so the model learns the correspondences between a gene, its transcript, and its protein product; the third stage exposes the model to protein structure data to sharpen its representation of secondary structure. Downstream evaluation reported in the official repository covers both DNA and protein: 0.905 MCC on promoter detection, 0.894 MCC on splice-site prediction, 90.83% accuracy on protein secondary structure prediction with CD-GPT-1b-s, 75.8% accuracy on protein solubility, and 57.29 precision at L/5 on contact map prediction. On RNA-protein interaction it reaches 0.5224 MCC and 76.05% accuracy on RPI-369 and 0.8204 MCC with 90.8% accuracy on RPI-488. The published checkpoints omit task-specific output-head weights, so downstream heads must be re-fit.
CD-GPT is aimed at multi-omic analyses where the interesting question crosses molecule boundaries: annotating regulatory elements in a genome, predicting properties of the proteins those genes encode, screening candidate RNA-protein binding pairs, or designing codon-optimized coding sequences for a target protein. Because one model covers all of these, groups working across transcriptomics and proteomics can standardize on a single embedding space instead of maintaining separate DNA and protein pipelines.
CD-GPT is one of the clearest tests of whether the central dogma is a useful inductive bias for sequence modeling, and its cross-molecular results support the case that jointly trained representations transfer where separately trained ones do not. Practical adoption has limits worth naming: the weights are distributed through Google Drive and Tencent Weiyun rather than a model hub, the repository carries a PolyForm Noncommercial 1.0.0 license that rules out commercial use, and the preprint has not been peer reviewed. The design has nonetheless informed later work on unified biological sequence models that treat nucleotides and amino acids as one language.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.